openinference
OpenInference TypeSafe AI Instrumentation
Python auto-instrumentation library for the TypeSafe AI Python SDK (typesafe-sdk).
Calls to TypeSafeClient.system_one and AsyncTypeSafeClient.system_one are traced and exported as OpenInference DECISION spans. A decision model scores or selects among the candidate options supplied in the request instead of generating free-form text, which is exactly what System One does: a request sends a state plus a map of typed questions (Noul, Choice, Score) and returns one typed answer per question, so the span records:
input.value: the request body (state,model,questions) as JSONllm.invocation_parameters: the call configuration, meaning themodeland anyextra_bodyfieldsoutput.value: the response body (model,answers,usage) as JSONdecision.systemanddecision.provider: bothtypesafedecision.request.model_name(for examplejev-latest),decision.response.model_name(the resolved model, for examplejev-1.13.0), anddecision.model_name(the response model when reported, else the requested one)decision.token_count.inputanddecision.token_count.output, from the response’susage
The model and its token usage are recorded under decision.* rather than llm.*, following the OpenInference decision span conventions, so decision model usage is not priced or counted as LLM usage.
A System One call is not a chat exchange, so the state and the answers are recorded only as input.value and output.value, not as llm.input_messages / llm.output_messages.
The state and the questions are recorded only in input.value, so TraceConfig(hide_inputs=True) keeps the whole request — caller data and question instructions alike — off the span, and hide_outputs=True does the same for the answers. llm.invocation_parameters holds no request content, only the model and any extra_body fields; mask it with hide_llm_invocation_parameters if those are sensitive.
These traces are fully OpenTelemetry compatible and can be sent to an OpenTelemetry collector for viewing, such as Arize Phoenix or Arize AX.
Supported Features
- Synchronous and asynchronous clients (
TypeSafeClientandAsyncTypeSafeClient) - All three question primitives, passed as SDK objects or raw dictionaries
- Suppressing tracing via
suppress_tracing() - Context attribute propagation (
using_session,using_user,using_attributes, metadata, tags) - Sensitive-data masking via
TraceConfig(e.g.hide_inputs,hide_outputs)
Requires typesafe-sdk >= 0.6.0.
Installation
pip install openinference-instrumentation-typesafe
Quickstart
pip install openinference-instrumentation-typesafe typesafe-sdk arize-phoenix opentelemetry-sdk opentelemetry-exporter-otlp
Start Phoenix as a collector (default http://localhost:6006), then:
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from openinference.instrumentation.typesafe import TypeSafeAIInstrumentor
endpoint = "http://127.0.0.1:6006/v1/traces"
tracer_provider = TracerProvider()
tracer_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter(endpoint)))
TypeSafeAIInstrumentor().instrument(tracer_provider=tracer_provider)
Make a System One call. Set the TYPESAFE_API_KEY environment variable with your key.
from typesafe_sdk import Choice, Noul, Score, TypeSafeClient
client = TypeSafeClient()
response = client.system_one(
state={"document": "I was charged twice. Please fix this ASAP."},
questions={
"billing": Noul(instructions="Is this ticket about billing?"),
"tone": Choice(
instructions="What is the customer's tone?",
criteria={"calm": None, "frustrated": None, "angry": None},
),
"urgency": Score(
instructions="How urgent is this ticket?",
criteria=["can wait", "this week", "today"],
),
},
)
print(response.nouls["billing"].noul)
print(response.choices["tone"].choice)
print(response.scores["urgency"].score)
Runnable examples, including async usage and context attributes, are in the examples/ directory.